Qwen Image Edit Two-Reference Element Swap Guide

Workflow Choice

Use the native Qwen-Image-Edit-2511 workflow when available

As of March 15, 2026, the official Qwen repository and the official ComfyUI docs both point to newer multi-image-capable Qwen edit workflows. The most practical choice for this task is Qwen-Image-Edit-2511, because it is the native ComfyUI workflow with improved consistency, and the official template explicitly exposes image, image2, and image3.

Why this is the best fit Your use case is not a generic single-image edit. It is a reference transfer task. Official Qwen materials say the 2509 line supports multiple image inputs, and the official 2511 workflow keeps that multi-image pattern while improving consistency.

Recommended path

Open the built-in Qwen-Image-Edit-2511 template in ComfyUI.
Use image as the base image you want to keep.
Use image2 as the donor image containing the element or material to borrow.

If 2511 is missing

Update ComfyUI first.
If your stable build still lacks it, use the official Qwen-Image-Edit-2509 multi-image workflow as fallback.
Expect slightly more drift and more prompt iteration.

Files and Placement

Put the required model files in the exact folders ComfyUI expects

The native workflow is easiest when the core files are already in their standard locations. The official ComfyUI docs for Qwen-Image-Edit-2511 list the following model set.

File	Purpose	Folder
`qwen_image_edit_2511_bf16.safetensors`	Main diffusion model	`ComfyUI/models/diffusion_models`
`qwen_2.5_vl_7b_fp8_scaled.safetensors`	Text / vision encoder	`ComfyUI/models/text_encoders`
`qwen_image_vae.safetensors`	VAE	`ComfyUI/models/vae`
`Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors`	Optional speed-up LoRA	`ComfyUI/models/loras`

Before you continue Launch ComfyUI once after copying the files so the model pickers refresh. If a model does not appear in the dropdown, a restart is usually the first fix.

Folder placement map

The first three files are the core setup. The Lightning LoRA is optional and belongs in loras, not in the model folders above.

Mental Model

Treat the two images as base and donor, not as equal references

The clearest way to think about this workflow is: image 1 is the picture you mostly want to preserve, and image 2 is the source of the specific thing you want to transfer. That framing should also appear in your prompt.

Image 1

Base scene, base subject, base composition, base lighting direction when possible.

Image 2

Donor element, donor texture, donor object, donor clothing detail, or donor accessory.

Prompt

Explicitly say what to preserve from image 1 and what to borrow from image 2.

Goal

Minimal unintended changes outside the swapped element.

Common mistake If you describe both references vaguely, Qwen may blend them loosely instead of performing a controlled transfer. Name the role of each image directly: base image versus donor image.

Step-by-Step Setup

Build the workflow around the native two-image edit node

Update ComfyUI and open the built-in Qwen-Image-Edit-2511 template. In official docs, missing templates usually mean the build is outdated.
Confirm the workflow’s model selectors point to qwen_image_edit_2511_bf16.safetensors, qwen_2.5_vl_7b_fp8_scaled.safetensors, and qwen_image_vae.safetensors.
Load your destination image into the first image input. This is the composition you want to keep.
Load your donor image into the second image input. This should show the element, texture, or object you want to transfer.
Enter a prompt that explicitly maps the edit: preserve image 1, borrow one named attribute from image 2, avoid unrelated scene changes.
Run one clean test before making the prompt more complex. Start simple and descriptive rather than cinematic.

Expected result after the first good pass The overall framing, camera angle, and scene layout should still look like image 1, while the swapped feature should visibly reflect image 2.

Best base image

Clear composition. One obvious subject. The region you want to modify is easy to identify.

Best donor image

The borrowed element is large, readable, and visually distinct from the background.

Best first test

Swap a material, accessory, logo area, or clothing feature before attempting complex full-subject fusion.

Workflow anatomy for a two-reference element swap

1. Load the base

Image 1 should be the image whose framing, pose, and scene you want to preserve.

2. Load the donor

Image 2 should isolate the thing you want the model to borrow as clearly as possible.

3. Constrain the output

The prompt should define what stays, what swaps, and what must not change.

Prompting Pattern

Use prompts that assign roles and limit the scope of the edit

The official Qwen examples for multi-image editing explicitly refer to the input images by role, and the official ComfyUI 2511 template ships with a donor-transfer example. For element swaps, that same pattern is the clearest way to prompt.

Keep the composition, subject placement, and overall scene from image 1.
Replace only the [element] in image 1 with the [element] from image 2.
Match the transferred element to the perspective and lighting of image 1.
Do not change the background, pose, camera angle, or any unrelated parts of image 1.

Good swap prompts

Change the jacket material in image 1 to the fur texture from image 2.
Keep the face and pose from image 1, but replace the sunglasses with the glasses from image 2.
Preserve image 1 as the base scene and move the flower crown from image 2 onto the subject in image 1.

Prompts to avoid

Mix these two images together.
Use image 2 style somehow.
Make image 1 like image 2 but also keep image 1.

Prompt-writing rule First say what stays. Then say what changes. Then say what should not change. That order usually gives a more stable edit than jumping straight to the replacement request.

Result Check

Judge the output against three simple questions

Did it keep image 1?

Composition, subject placement, pose, and environment should still read as the base image.

Did it use image 2?

The transferred element should clearly resemble the donor image, not a generic substitute.

Did it stay local?

The edit should not rewrite the whole image unless your prompt asked for a broader transformation.

Good first-run target Aim for “recognizably correct and mostly stable,” not perfection. Once the model is reading both references correctly, refinement becomes much easier.

Troubleshooting

Fix the likely failure mode nearest to the symptom

Problem: it ignores image 2 Make the donor element more explicit in the prompt, and describe it by name. If needed, use a cleaner donor image or crop the donor image so the desired element is larger and easier to read.

Problem: it changes too much of image 1 Add stronger preservation language such as “keep the composition, pose, background, and camera angle from image 1” and reduce the request to one local change.

Problem: models are missing in dropdowns Check the exact filenames and folders, then restart ComfyUI. The official docs use standard core folders, so placement mistakes are usually the cause.

Problem: the workflow template is missing Update ComfyUI. Official documentation notes that missing native Qwen templates often mean the build is not current enough yet.

Practical refinement order Improve the prompt first, then improve the donor image clarity, then try the fallback 2509 workflow only if your build cannot use 2511.

Research Notes

Official sources used for this guide

The guidance above is based on current official ComfyUI and Qwen sources, plus the official Comfy workflow template JSON.

Key detail confirmed from the official template The native Qwen-Image-Edit-2511 workflow exposes separate inputs for image and image2, and the bundled example prompt performs the same kind of donor-to-base transfer this guide is built around.